HFTG Feature Extraction for Metaheuristic HW/SW-Partitioning：An H.264 encoder case study

• Main Authors: Wen-Hao Chang, 張文豪
• Other Authors: Ruei-xi Chen
• Format: Others
• Language: zh-TW
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/85593873064635690482

碩士 === 聖約翰科技大學 === 電機工程系碩士班 === 99 === The decision accuracy of hardware/software partitioning is now playing a key role towards success for an embedded system design. Recently, meta-heuristic algorithms such as GA or PSO are often exploited for solving the decision problems. Unfortunately, the use of single fitness value for the algorithm is difficult to present the design quality that concerned the factors of performance, cost, power consumption, communication latency and chip area, etc. So that bad decisions are often obtained. In this paper, we proposed a novel Hot Function Task Group (HFTG) feature extract algorithm to improve the decision efficiency of meta-heuristic algorithms. The goal is originally to avoid the trap of decision and then reduce the complexity of design space. First, we explore the call-graph structure for the functions in a system. Next, we modify fitness function based-on the cost of interface communications to extract groups. Then delete the decision points that have high communication cost to them. Finally, the control-data-flow-graph (CDFG) of the rest functions and the features are delivered to the meta-heuristic algorithms for hardware-software partitioning. For testing the efficiency of the algorithm, a JM-code of H.264 encoder, the advanced video coding standard, is employed as the study case. While extracting 79 functions from 163 functions of H.264, it results in obtaining six groups of HTFGs which occupying more than half of the system execution time. It revealed that HTFG feature extraction method chooses the blocks with high complexity and relatively simple interfaces for accelerating and avoids un-proper hardware-software calling style. The HTFG method that can do correct partition and implementation for functions in a system is proved to be valuable in the field of hardware/software co-design.